Assessing the Environmental and Human Cost of AI

The advent of large language models (LLMs) has revolutionized the fields of artificial intelligence, machine learning, and natural language processing. LLMs, which include tools like OpenAI's ChatGPT, Google's Bard, and others, have demonstrated remarkable capabilities, from answering complex questions to assisting in creative endeavors. However, training and maintaining these models require substantial computational resources, translating into significant environmental impacts. Chief among these impacts is water usage, primarily to cool the vast data centers that run these models, raising the question: are the benefits provided by LLMs worth the strain on the world's increasingly scarce fresh water supplies?

water footprint

Water Consumption and LLM Data Centers

Large data centers are intensive consumers of fresh water due to the cooling needs of high-powered processors. Each data center houses thousands of servers, each of which generates considerable heat. Cooling systems, often involving water-based cooling towers, dissipate this heat to maintain optimal operating temperatures. Data centers in the U.S. alone use billions of liters of water annually. For example, a report from the Environmental Protection Agency highlights that data centers are among the top water consumers within industrial settings, using anywhere from 1.7 to 3.6 liters per kWh for cooling (EPA). With the expansion of LLMs, many of these centers have grown to accommodate more powerful servers, leading to even higher water usage.

In a 2023 study, researchers at the University of California estimated that Google's data centers required 15.79 billion liters of water in 2021 alone (UC Berkeley). Microsoft's reported data center consumption spiked with its AI model expansion, using 6.8 billion liters of water in 2022, nearly doubling from previous years (Microsoft). This growth trend suggests a grim outlook for the future if AI and machine learning development continues to intensify.

The Challenge of Saltwater Desalination

Given the strain on fresh water resources, some argue that turning to saltwater, which makes up approximately 97% of the world's water, is a potential solution. Desalination is a process that removes minerals and salts from seawater, making it fit for human consumption. Yet, desalination technology is highly energy-intensive, requiring approximately 3-10 kWh of electricity to produce one cubic meter of fresh water (UCS). This energy usage contributes to carbon emissions, making desalination a less attractive option for mitigating the environmental impact of fresh water use in data centers.

A widely used method for desalination is reverse osmosis, which forces water through a semi-permeable membrane to filter out salts. While effective, it is not without drawbacks: the high pressure required for the process consumes vast amounts of energy, and only about 50% of the seawater can be converted into fresh water, with the remaining brine waste requiring careful disposal to prevent ecological harm (NOAA).

Energy Requirements for Desalination and Its Environmental Impact

The power consumption required to desalinate seawater is substantial. According to a study published in the journal Nature Sustainability, desalinating 1000 cubic meters of seawater requires roughly 5 megawatt-hours (MWh) of electricity (Nature Sustainability). To put this in perspective, maintaining fresh water for an LLM's data center could mean building a large-scale desalination plant to ensure sustainable water access without depleting local freshwater resources.

But desalination isn't a “clean” solution; it also contributes to climate change. The energy input for desalination largely depends on fossil fuels, with renewable sources making up a small percentage of power in current desalination facilities. The cumulative effect on greenhouse gas emissions adds another layer of environmental cost, which could be exacerbated if desalination becomes a primary means of supporting water-intensive data centers.

Fresh Water, Food Supply, and the Social Impact of LLMs

The demand for fresh water extends beyond data centers to crucial areas like agriculture, where water is essential for crop production. As global population grows, food production will need to increase by roughly 60% by 2050, according to the Food and Agriculture Organization (FAO). Agriculture currently uses around 70% of global freshwater resources, making it highly susceptible to any diversion of water toward non-agricultural industries like data centers. For communities facing water scarcity, competition for fresh water exacerbates food insecurity, impacting millions.

The implications of allocating vast water resources to LLMs, whose responses are often inaccurate, become stark in this context. Despite recent advancements, LLMs can still generate incorrect or misleading information, requiring human oversight for high-stakes applications. For instance, LLMs often fail to distinguish reliable from unreliable sources, leading to factual errors in medical, legal, and scientific information. This inaccuracy necessitates questioning whether the environmental cost of LLMs outweighs their benefits.

Is the Fresh Water Trade-Off Worth It?

The rapid development and deployment of LLMs have driven many companies to prioritize technological advancement over environmental sustainability. While LLMs have certainly enhanced productivity and transformed several industries, the question remains whether their benefits justify the environmental cost — particularly when their information reliability is inconsistent. In a world where water scarcity is growing, and agriculture and drinking supplies rely heavily on freshwater availability, the expansion of LLM infrastructure could detract from human and environmental well-being.

Some tech companies, including Google and Microsoft, are making strides in using recycled water and enhancing cooling efficiencies. However, without a substantial change in data center water practices or the increased use of alternative cooling systems, the water usage trend will likely continue to rise. Additionally, integrating LLMs in society and relying on them in vital roles (e.g., educational aids, business consultants, and research tools) could increase pressure on freshwater resources even more if the energy and water demands are not addressed.

A Call for Sustainable AI Development

LLMs hold transformative potential for enhancing human knowledge and productivity, but their environmental footprint is significant. Given the high water usage of data centers running LLMs and the challenges of desalination, the viability of LLMs on a global scale must be carefully considered against the needs of water-dependent systems like agriculture. Moreover, as long as LLMs remain prone to inaccuracies, society should question whether their benefits outweigh the environmental impact and potential threat to food security.

The development of sustainable AI practices, such as innovating alternative cooling technologies, increasing the use of recycled water, and improving the efficiency of desalination plants, is essential. Without these changes, the continued growth of AI-powered applications may come at too high a cost, especially for the millions who rely on fresh water for basic survival and food security. As such, policymakers, environmentalists, and the tech industry must work together to ensure that the progress of AI does not lead to environmental regression.

References:

Environmental Protection Agency. (n.d.). Data centers and energy. https://www.epa.gov/datacenters/data-centers-energy

Food and Agriculture Organization. (n.d.). Food security and nutrition: Building a climate-resilient future. https://www.fao.org/publications/sofi/2021/en/

Microsoft. (n.d.). Environmental sustainability in our datacenters. https://www.microsoft.com/en-us/sustainability/emissions-impact-datacenters

National Oceanic and Atmospheric Administration. (n.d.). Desalination and water purification research program. https://www.noaa.gov/education/resource-collections/freshwater/desalination

Nature Sustainability. (2023). Environmental impact of desalination processes. https://www.nature.com/articles/s41893-023-01113-9

University of California, Berkeley. (2023). Water use in data centers study. https://news.berkeley.edu/2023/03/13/how-much-water-do-data-centers-use

Union of Concerned Scientists. (n.d.). The energy-water nexus: How energy and water supplies are linked. https://www.ucsusa.org/resources/energy-and-water

water (1) ai (1) llm (1)